solow-cv
Cross-validation and resampling for Rust. Deterministic K-fold, stratified, group-aware, time-series, and leakage-safe purged splitters. Learning / validation / permutation curves. Bootstrap confidence intervals. Time-series block bootstrap. Optional rayon-parallel fold evaluation.
Part of the Solow statistics and machine learning stack.
Install
[dependencies]
solow-cv = "0.7"
Quick example
use solow_cv::{StratifiedKFold, cross_val_score};
let cv = StratifiedKFold::new(5)?.shuffle(true).seed(42);
let scores = cross_val_score(&cv, n, |train, test| {
let m = fit_your_model(&x, &y, train)?;
Ok(m.score(&x, &y, test)?)
})?;
println!("{:.4} +/- {:.4}", scores.mean, scores.std);
What is inside
Random and stratified splitters
| Splitter |
What it does |
KFold |
K-fold with optional shuffle and seed. Standard "first n mod k folds are one bigger" convention. |
StratifiedKFold |
K-fold preserving per-fold class prevalence. |
RepeatedKFold, RepeatedStratifiedKFold |
Multiple K-fold repetitions with different seeds. |
ShuffleSplit, StratifiedShuffleSplit |
Random n-times splits with configurable test size. |
LeaveOneOut |
n folds of size 1. |
LeavePOut |
All C(n, p) folds of size p. |
Group-aware splitters
| Splitter |
What it does |
GroupKFold |
Every group in exactly one fold, via greedy longest-first balancer. |
StratifiedGroupKFold |
GroupKFold that also preserves class prevalence via a round-robin seed and SSE minimisation. |
GroupShuffleSplit |
Group-aware random splits. |
Time series
| Splitter |
What it does |
TimeSeriesSplit |
Walk-forward validation with configurable test_size, gap, and max_train_size for rolling windows. |
PurgedKFold |
López de Prado (2018) walk-forward K-fold with symmetric purge band around every test block and optional trailing embargo. |
CombinatorialPurgedKFold |
CPCV that enumerates every C(n_splits, n_test) subset of contiguous blocks, giving a much richer set of distinct back-test paths than plain K-fold. |
Scoring and curves
| Function |
What it does |
cross_val_score |
Run any fit-and-score callback across every fold. Returns per-fold scores plus mean, unbiased standard deviation, and standard error. |
cross_val_score_parallel |
Same, but evaluates folds concurrently on a rayon thread pool (opt-in parallel feature). |
learning_curve |
Training-size sweep for diagnosing bias / variance. |
validation_curve |
Hyperparameter sweep for regularization curves. |
permutation_test_score |
Non-parametric significance test for a cross-val score. |
Bootstrap
| Function |
What it does |
bootstrap_ci |
Percentile, basic ("reverse-percentile"), and Efron BCa confidence intervals via jackknife acceleration. All three computed from the same replicates. |
moving_block_bootstrap_indices |
Overlapping blocks of fixed length for time-series resampling. |
circular_block_bootstrap_indices |
Wrap-around blocks that avoid the end-of-series bias. |
stationary_bootstrap_indices |
Politis-Romano stationary bootstrap with geometric block lengths. |
Determinism
Random splits and bootstrap resamples use a seeded MMIX-constants 64-bit LCG. Folds and replicates are bit-for-bit reproducible across runs and platforms.
serde and parallel features
[dependencies]
solow-cv = { version = "0.7", features = ["serde", "parallel"] }
Correctness
- Every splitter is deterministic under a fixed seed.
- Cross-verified against committed golden reference fixtures.
#![forbid(unsafe_code)].
License
BSD-3-Clause. See the Solow workspace for the full stack.